Research Of Automatic Image Annotation And Image Retrieval Algorithm

With the rapid development of mobile Internet, billions of images will be produced everyday. Since these images are organized in mess, in order to deal with and process these massive images efficiently and retrieve the useful images, the image retrieval techniques are proposed. The content-based image retrieval and the semantic-based image retrieval are the most widely used algorithms for the image retrieval problem. And automatic image annotation is an important solution of obtaining semantics. However, during the process of automatic image annotation, there still exist some problems need to be further solved.1. The "semantic gap" problem between the low visual features and high semantic features of an image; 2. the "class-imbalance" problem that inspired by high variance in the number of images corresponding to different labels; 3. the "incomplete-labeling" problem that inspired by incompletely tagged images in the training image sets. This thesis mainly concentrates on solving the above-mentioned problem. The main work of this thesis is conducted as follows:To alleviate the semantic gap, class-imbalance, and incomplete-labeling problems, an automatic image annotation algorithm based on canonical correlation analytical subspace and K-nearest neighbor is proposed. First, the low level visual features and the high level semantic features are projected to a common feature subspace by a CCA method. The correlations between the low level visual features and the high level semantic features are setup in the common feature subspace. Then, the relevance between images and labels are obtained according to the correlation between low visual features and high semantic features. And the semantic space of each label is further defined. The semantic space of each label is made up by the image subset in the CCA subspace. And then, a K-nearest neighbor algorithm is used to obtain the semantic neighbor in the semantic space of each label, so that the scale of the image subset of each label can be balanced. This image subset contains almost all of the labels from vocabulary, therefore, the labeling probability between the labels from vocabulary and the unlabeled image can be obtained. In the annotating process, a Bayes probability model is constructed by using the visual distance between semantic subsets and the unlabeled images, with which the annotation task can be performed. Finally, the correlation between labels is used to improve the image annotation quality. The comparative experimental results on the Corel5k, ESP Game, IAPR TC-12 benchmark image set illustrate that the proposed algorithm can effectively perform image annotation tasks.To improve the retrieval accuracy of the hash-based image retrieval algorithm, an image retrieval algorithm based on block hashing is proposed. First, the high-dimensional feature vector is divided into blocks and each block is projected separately using specific hashing methods, then all the real-valued vectors generated from the projection stage are combined into a long vector. And then, the long vector is quantized into binary codes by thresholding. For the training set and the querying set, the above process is repeated separately to obtain the corresponding hash code. Finally, the Hamming distance is adopted to measure the similarity between points in the hash code space. For a given query image, several retrieval images which have the minimum Hamming distance are assigned. The performance of the proposed algorithm is tested on Caltech-256 and CIFAR-10 dataset. The results show that the proposed algorithm can obtain higher accuracy with requiring less query time than some other state of the art image retrieval algorithms.